PTX: Adjust rounding modes

* rounding modes for fp add, mul, sub now use .rn
* float -> int rounding correctly uses .rzi not .rni
* 32bit fdiv for sm13 uses div.rn (instead of div.approx)
* 32bit fdiv for sm10 now uses div (instead of div.approx)

Approx is not IEEE 754 compatible (and should be optionally set by a flag to the backend instead). The .rn rounding modifier is the PTX default anyway, but it's better to be explicit.

All these modifiers should be available by using __fmul_rz functions for example, but support will need to be added for this in the backend.

Patch by Dan Bailey

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@133253 91177308-0d34-0410-b5e6-96231b3b80d8

7 files changed, 60 insertions, 45 deletions

diff --git a/lib/Target/PTX/PTXInstrInfo.td b/lib/Target/PTX/PTXInstrInfo.td
index 71f7cc32b8..8477cd71d3 100644
--- a/lib/Target/PTX/PTXInstrInfo.td
+++ b/lib/Target/PTX/PTXInstrInfo.td
@@ -584,24 +584,39 @@ defm REM : INT3<"rem", urem>;
 defm FNEG : PTX_FLOAT_2OP<"neg", fneg>;
 
 // Standard Binary Operations
-defm FADD : PTX_FLOAT_3OP<"add", fadd>;
-defm FSUB : PTX_FLOAT_3OP<"sub", fsub>;
-defm FMUL : PTX_FLOAT_3OP<"mul", fmul>;
-
-// TODO: Allow user selection of rounding modes for fdiv.
-// For division, we need to have f32 and f64 differently.
-// For f32, we just always use .approx since it is supported on all hardware
-// for PTX 1.4+, which is our minimum target.
-def FDIVrr32 : InstPTX<(outs RegF32:$d),
+defm FADD : PTX_FLOAT_3OP<"add.rn", fadd>;
+defm FSUB : PTX_FLOAT_3OP<"sub.rn", fsub>;
+defm FMUL : PTX_FLOAT_3OP<"mul.rn", fmul>;
+
+// For floating-point division:
+// SM_13+ defaults to .rn for f32 and f64,
+// SM10 must *not* provide a rounding
+
+// TODO: 
+//     - Allow user selection of rounding modes for fdiv
+//     - Add support for -prec-div=false (.approx)
+
+def FDIVrr32SM13 : InstPTX<(outs RegF32:$d),
                        (ins RegF32:$a, RegF32:$b),
-                       "div.approx.f32\t$d, $a, $b",
-                       [(set RegF32:$d, (fdiv RegF32:$a, RegF32:$b))]>;
-def FDIVri32 : InstPTX<(outs RegF32:$d),
+                       "div.rn.f32\t$d, $a, $b",
+                       [(set RegF32:$d, (fdiv RegF32:$a, RegF32:$b))]>,
+                   Requires<[SupportsSM13]>;
+def FDIVri32SM13 : InstPTX<(outs RegF32:$d),
                        (ins RegF32:$a, f32imm:$b),
-                       "div.approx.f32\t$d, $a, $b",
-                       [(set RegF32:$d, (fdiv RegF32:$a, fpimm:$b))]>;
+                       "div.rn.f32\t$d, $a, $b",
+                       [(set RegF32:$d, (fdiv RegF32:$a, fpimm:$b))]>,
+                   Requires<[SupportsSM13]>;
+def FDIVrr32SM10 : InstPTX<(outs RegF32:$d),
+                       (ins RegF32:$a, RegF32:$b),
+                       "div.f32\t$d, $a, $b",
+                       [(set RegF32:$d, (fdiv RegF32:$a, RegF32:$b))]>,
+                   Requires<[DoesNotSupportSM13]>;
+def FDIVri32SM10 : InstPTX<(outs RegF32:$d),
+                       (ins RegF32:$a, f32imm:$b),
+                       "div.f32\t$d, $a, $b",
+                       [(set RegF32:$d, (fdiv RegF32:$a, fpimm:$b))]>,
+                   Requires<[DoesNotSupportSM13]>;
 
-// For f64, we must specify a rounding for sm 1.3+ but *not* for sm 1.0.
 def FDIVrr64SM13 : InstPTX<(outs RegF64:$d),
                            (ins RegF64:$a, RegF64:$b),
                            "div.rn.f64\t$d, $a, $b",
@@ -825,11 +840,11 @@ def CVT_pred_u64
             [(set RegPred:$d, (trunc RegI64:$a))]>;
 
 def CVT_pred_f32
-  : InstPTX<(outs RegPred:$d), (ins RegF32:$a), "cvt.rni.pred.f32\t$d, $a",
+  : InstPTX<(outs RegPred:$d), (ins RegF32:$a), "cvt.rzi.pred.f32\t$d, $a",
             [(set RegPred:$d, (fp_to_uint RegF32:$a))]>;
 
 def CVT_pred_f64
-  : InstPTX<(outs RegPred:$d), (ins RegF64:$a), "cvt.rni.pred.f64\t$d, $a",
+  : InstPTX<(outs RegPred:$d), (ins RegF64:$a), "cvt.rzi.pred.f64\t$d, $a",
             [(set RegPred:$d, (fp_to_uint RegF64:$a))]>;
 
 // Conversion to u16
@@ -847,11 +862,11 @@ def CVT_u16_u64
             [(set RegI16:$d, (trunc RegI64:$a))]>;
 
 def CVT_u16_f32
-  : InstPTX<(outs RegI16:$d), (ins RegF32:$a), "cvt.rni.u16.f32\t$d, $a",
+  : InstPTX<(outs RegI16:$d), (ins RegF32:$a), "cvt.rzi.u16.f32\t$d, $a",
             [(set RegI16:$d, (fp_to_uint RegF32:$a))]>;
 
 def CVT_u16_f64
-  : InstPTX<(outs RegI16:$d), (ins RegF64:$a), "cvt.rni.u16.f64\t$d, $a",
+  : InstPTX<(outs RegI16:$d), (ins RegF64:$a), "cvt.rzi.u16.f64\t$d, $a",
             [(set RegI16:$d, (fp_to_uint RegF64:$a))]>;
 
 // Conversion to u32
@@ -869,11 +884,11 @@ def CVT_u32_u64
             [(set RegI32:$d, (trunc RegI64:$a))]>;
 
 def CVT_u32_f32
-  : InstPTX<(outs RegI32:$d), (ins RegF32:$a), "cvt.rni.u32.f32\t$d, $a",
+  : InstPTX<(outs RegI32:$d), (ins RegF32:$a), "cvt.rzi.u32.f32\t$d, $a",
             [(set RegI32:$d, (fp_to_uint RegF32:$a))]>;
 
 def CVT_u32_f64
-  : InstPTX<(outs RegI32:$d), (ins RegF64:$a), "cvt.rni.u32.f64\t$d, $a",
+  : InstPTX<(outs RegI32:$d), (ins RegF64:$a), "cvt.rzi.u32.f64\t$d, $a",
             [(set RegI32:$d, (fp_to_uint RegF64:$a))]>;
 
 // Conversion to u64
@@ -891,11 +906,11 @@ def CVT_u64_u32
             [(set RegI64:$d, (zext RegI32:$a))]>;
 
 def CVT_u64_f32
-  : InstPTX<(outs RegI64:$d), (ins RegF32:$a), "cvt.rni.u64.f32\t$d, $a",
+  : InstPTX<(outs RegI64:$d), (ins RegF32:$a), "cvt.rzi.u64.f32\t$d, $a",
             [(set RegI64:$d, (fp_to_uint RegF32:$a))]>;
 
 def CVT_u64_f64
-  : InstPTX<(outs RegI64:$d), (ins RegF64:$a), "cvt.rni.u64.f64\t$d, $a",
+  : InstPTX<(outs RegI64:$d), (ins RegF64:$a), "cvt.rzi.u64.f64\t$d, $a",
             [(set RegI64:$d, (fp_to_uint RegF64:$a))]>;
 
 // Conversion to f32
diff --git a/test/CodeGen/PTX/add.ll b/test/CodeGen/PTX/add.ll
index b89a2f6269..c16be4933c 100644
--- a/test/CodeGen/PTX/add.ll
+++ b/test/CodeGen/PTX/add.ll
@@ -22,14 +22,14 @@ define ptx_device i64 @t1_u64(i64 %x, i64 %y) {
 }
 
 define ptx_device float @t1_f32(float %x, float %y) {
-; CHECK: add.f32 r0, r1, r2
+; CHECK: add.rn.f32 r0, r1, r2
 ; CHECK-NEXT: ret;
   %z = fadd float %x, %y
   ret float %z
 }
 
 define ptx_device double @t1_f64(double %x, double %y) {
-; CHECK: add.f64 rd0, rd1, rd2
+; CHECK: add.rn.f64 rd0, rd1, rd2
 ; CHECK-NEXT: ret;
   %z = fadd double %x, %y
   ret double %z
@@ -57,14 +57,14 @@ define ptx_device i64 @t2_u64(i64 %x) {
 }
 
 define ptx_device float @t2_f32(float %x) {
-; CHECK: add.f32 r0, r1, 0F3F800000;
+; CHECK: add.rn.f32 r0, r1, 0F3F800000;
 ; CHECK-NEXT: ret;
   %z = fadd float %x, 1.0
   ret float %z
 }
 
 define ptx_device double @t2_f64(double %x) {
-; CHECK: add.f64 rd0, rd1, 0D3FF0000000000000;
+; CHECK: add.rn.f64 rd0, rd1, 0D3FF0000000000000;
 ; CHECK-NEXT: ret;
   %z = fadd double %x, 1.0
   ret double %z
diff --git a/test/CodeGen/PTX/cvt.ll b/test/CodeGen/PTX/cvt.ll
index 984cb4d8d5..bf18bd7921 100644
--- a/test/CodeGen/PTX/cvt.ll
+++ b/test/CodeGen/PTX/cvt.ll
@@ -31,7 +31,7 @@ define ptx_device i32 @cvt_pred_i64(i64 %x, i1 %y) {
 }
 
 define ptx_device i32 @cvt_pred_f32(float %x, i1 %y) {
-; CHECK: cvt.rni.pred.f32 p0, r1;
+; CHECK: cvt.rzi.pred.f32 p0, r1;
 ; CHECK: ret;
 	%a = fptoui float %x to i1
 	%b = and i1 %a, %y
@@ -40,7 +40,7 @@ define ptx_device i32 @cvt_pred_f32(float %x, i1 %y) {
 }
 
 define ptx_device i32 @cvt_pred_f64(double %x, i1 %y) {
-; CHECK: cvt.rni.pred.f64 p0, rd1;
+; CHECK: cvt.rzi.pred.f64 p0, rd1;
 ; CHECK: ret;
 	%a = fptoui double %x to i1
 	%b = and i1 %a, %y
@@ -72,14 +72,14 @@ define ptx_device i16 @cvt_i16_i64(i64 %x) {
 }
 
 define ptx_device i16 @cvt_i16_f32(float %x) {
-; CHECK: cvt.rni.u16.f32 rh0, r1;
+; CHECK: cvt.rzi.u16.f32 rh0, r1;
 ; CHECK: ret;
 	%a = fptoui float %x to i16
 	ret i16 %a
 }
 
 define ptx_device i16 @cvt_i16_f64(double %x) {
-; CHECK: cvt.rni.u16.f64 rh0, rd1;
+; CHECK: cvt.rzi.u16.f64 rh0, rd1;
 ; CHECK: ret;
 	%a = fptoui double %x to i16
 	ret i16 %a
@@ -109,14 +109,14 @@ define ptx_device i32 @cvt_i32_i64(i64 %x) {
 }
 
 define ptx_device i32 @cvt_i32_f32(float %x) {
-; CHECK: cvt.rni.u32.f32 r0, r1;
+; CHECK: cvt.rzi.u32.f32 r0, r1;
 ; CHECK: ret;
 	%a = fptoui float %x to i32
 	ret i32 %a
 }
 
 define ptx_device i32 @cvt_i32_f64(double %x) {
-; CHECK: cvt.rni.u32.f64 r0, rd1;
+; CHECK: cvt.rzi.u32.f64 r0, rd1;
 ; CHECK: ret;
 	%a = fptoui double %x to i32
 	ret i32 %a
@@ -146,14 +146,14 @@ define ptx_device i64 @cvt_i64_i32(i32 %x) {
 }
 
 define ptx_device i64 @cvt_i64_f32(float %x) {
-; CHECK: cvt.rni.u64.f32 rd0, r1;
+; CHECK: cvt.rzi.u64.f32 rd0, r1;
 ; CHECK: ret;
 	%a = fptoui float %x to i64
 	ret i64 %a
 }
 
 define ptx_device i64 @cvt_i64_f64(double %x) {
-; CHECK: cvt.rni.u64.f64 rd0, rd1;
+; CHECK: cvt.rzi.u64.f64 rd0, rd1;
 ; CHECK: ret;
 	%a = fptoui double %x to i64
 	ret i64 %a
diff --git a/test/CodeGen/PTX/fdiv-sm10.ll b/test/CodeGen/PTX/fdiv-sm10.ll
index 9aff25111b..eb32222f3a 100644
--- a/test/CodeGen/PTX/fdiv-sm10.ll
+++ b/test/CodeGen/PTX/fdiv-sm10.ll
@@ -1,7 +1,7 @@
 ; RUN: llc < %s -march=ptx32 -mattr=+sm10 | FileCheck %s
 
 define ptx_device float @t1_f32(float %x, float %y) {
-; CHECK: div.approx.f32 r0, r1, r2;
+; CHECK: div.f32 r0, r1, r2;
 ; CHECK-NEXT: ret;
 	%a = fdiv float %x, %y
 	ret float %a
diff --git a/test/CodeGen/PTX/fdiv-sm13.ll b/test/CodeGen/PTX/fdiv-sm13.ll
index 84e0adab7e..ad24f35b20 100644
--- a/test/CodeGen/PTX/fdiv-sm13.ll
+++ b/test/CodeGen/PTX/fdiv-sm13.ll
@@ -1,7 +1,7 @@
 ; RUN: llc < %s -march=ptx32 -mattr=+sm13 | FileCheck %s
 
 define ptx_device float @t1_f32(float %x, float %y) {
-; CHECK: div.approx.f32 r0, r1, r2;
+; CHECK: div.rn.f32 r0, r1, r2;
 ; CHECK-NEXT: ret;
 	%a = fdiv float %x, %y
 	ret float %a
diff --git a/test/CodeGen/PTX/mul.ll b/test/CodeGen/PTX/mul.ll
index 93f94e3509..2093556dac 100644
--- a/test/CodeGen/PTX/mul.ll
+++ b/test/CodeGen/PTX/mul.ll
@@ -11,28 +11,28 @@
 ;}
 
 define ptx_device float @t1_f32(float %x, float %y) {
-; CHECK: mul.f32 r0, r1, r2
+; CHECK: mul.rn.f32 r0, r1, r2
 ; CHECK-NEXT: ret;
   %z = fmul float %x, %y
   ret float %z
 }
 
 define ptx_device double @t1_f64(double %x, double %y) {
-; CHECK: mul.f64 rd0, rd1, rd2
+; CHECK: mul.rn.f64 rd0, rd1, rd2
 ; CHECK-NEXT: ret;
   %z = fmul double %x, %y
   ret double %z
 }
 
 define ptx_device float @t2_f32(float %x) {
-; CHECK: mul.f32 r0, r1, 0F40A00000;
+; CHECK: mul.rn.f32 r0, r1, 0F40A00000;
 ; CHECK-NEXT: ret;
   %z = fmul float %x, 5.0
   ret float %z
 }
 
 define ptx_device double @t2_f64(double %x) {
-; CHECK: mul.f64 rd0, rd1, 0D4014000000000000;
+; CHECK: mul.rn.f64 rd0, rd1, 0D4014000000000000;
 ; CHECK-NEXT: ret;
   %z = fmul double %x, 5.0
   ret double %z
diff --git a/test/CodeGen/PTX/sub.ll b/test/CodeGen/PTX/sub.ll
index 9efeaace0e..4d552801a4 100644
--- a/test/CodeGen/PTX/sub.ll
+++ b/test/CodeGen/PTX/sub.ll
@@ -22,14 +22,14 @@ define ptx_device i64 @t1_u64(i64 %x, i64 %y) {
 }
 
 define ptx_device float @t1_f32(float %x, float %y) {
-; CHECK: sub.f32 r0, r1, r2
+; CHECK: sub.rn.f32 r0, r1, r2
 ; CHECK-NEXT: ret;
   %z = fsub float %x, %y
   ret float %z
 }
 
 define ptx_device double @t1_f64(double %x, double %y) {
-; CHECK: sub.f64 rd0, rd1, rd2
+; CHECK: sub.rn.f64 rd0, rd1, rd2
 ; CHECK-NEXT: ret;
   %z = fsub double %x, %y
   ret double %z
@@ -57,14 +57,14 @@ define ptx_device i64 @t2_u64(i64 %x) {
 }
 
 define ptx_device float @t2_f32(float %x) {
-; CHECK: add.f32 r0, r1, 0FBF800000;
+; CHECK: add.rn.f32 r0, r1, 0FBF800000;
 ; CHECK-NEXT: ret;
   %z = fsub float %x, 1.0
   ret float %z
 }
 
 define ptx_device double @t2_f64(double %x) {
-; CHECK: add.f64 rd0, rd1, 0DBFF0000000000000;
+; CHECK: add.rn.f64 rd0, rd1, 0DBFF0000000000000;
 ; CHECK-NEXT: ret;
   %z = fsub double %x, 1.0
   ret double %z